Adam Johnson lives in West Virginia. To prepare for the US midterm elections, he did not discuss them with his friends or family. He spent hours talking about them with ChatGPT.
He first asked it to list every election on his ballot, then he dug into it, candidate by candidate. Today, he says he would feel ready to vote.
His story, reported by American public radio NPR, is not an isolated one. And it concerns us directly: France elects its president in six months.
What these voters are doing
In the United States, early voting has already begun. Several voters interviewed by NPR explain that they use an AI assistant in three ways.
For small local elections, the ones the press barely covers: a county council, a judgeship, a sheriff. An internet search returns few results; an assistant, by contrast, always answers.
To compare candidates, in table form. One voter in Wisconsin asked Gemini to compare two candidates for lieutenant governor.
To check a rumour circulating on social media.
The Boston Globe is already calling it the first AI election.
Where it breaks down
The answers are not reliable across the board. A study published in August by an American democracy watchdog found that ChatGPT and Google's AI gave incomplete and inconsistent answers to common voter questions.
The small-election paradox. It is precisely where information is scarce that voters use AI the most. But it is also where it is most likely to make things up: the fewer sources exist, the more it fills in the gaps. One voter even describes his workaround: checking one chatbot against another.
Invisible biases. The companies do not say which sources their models draw on. It is therefore difficult to know what shapes their answers. OpenAI says it monitors its model's biases to keep ChatGPT politically neutral. One voter sums up the question many are asking: who is paying to build these tools, and to what end?
Since a 2018 law against information manipulation, France has had a specific tool for election periods. During the three months preceding a national election, a judge can be petitioned on an urgent basis and must rule within 48 hours to stop the mass, artificial spread of false information online. The text was drafted in the era of fake accounts and sponsored content. It will apply to the 2027 presidential election, at a time when some voters will be getting their information from an AI.
Five habits for 2027
1. Ask for manifestos, not advice. "What does this candidate propose on pensions?" is a good question. "Who should I vote for?" is not. Assistants often refuse to answer the second, and rightly so.
2. Set your own criteria. Say what matters to you, and ask for a comparison on those points. You set the scorecard, not the tool.
3. Go back to the official text. Every candidate publishes their manifesto. If an AI summarises a proposal for you, go and read it at the source. We set out the method in detail in our guide to verifying a source.
4. Be even more wary with local elections. The less information is available, the greater the risk of error.
5. Cross-check with other sources. The press, debates, official websites. And be wary of content that looks too polished circulating during the campaign: AIs can already manufacture entire crowds of fake accounts.
What we take away
Using an AI to understand an election is not absurd. It can summarise a manifesto, explain a reform, organise a comparison. For many people, it is more accessible than a three-hour debate.
But understanding and choosing are two different things. An AI can help you understand. The choice itself must remain yours.